Kimi k0-math

Name: Kimi k0-math
Rating: 6

Moonshot's first reasoning model — an RL-scaled math specialist that rivaled OpenAI's o1-mini at launch.

Visit official page →

About

k0-math was Moonshot AI's opening move in the reasoning race, announced in November 2024 as the company's first model built to think rather than just predict tokens. Founder Yang Zhilin framed it as a shift in scaling philosophy: instead of bigger models, reinforcement learning that teaches an AI to explore mathematical problems deeply.

At launch it beat OpenAI's o1-mini and o1-preview on school-level, high-school, college-entrance and MATH benchmarks, and reached 90% of o1-mini's OMNI-MATH score and 83% on AIME. It rolled out inside the Kimi web platform and the Kimi smart assistant app, including a Kimi Exploration edition that applied its RL thinking to search. The model itself was never open-weighted, and Moonshot has since moved its flagship line to the K2 series.

Specs

Context window

Not published

Pricing

Free via the Kimi web platform and app; no standalone API pricing verified

Free quota

Free access on Kimi platform

API access

Available through the Kimi platform (web/app); no public weights

Weights

Closed — API only

Released

Not disclosed

Benchmarks

Moonshot-reported (Nov 2024): outperformed o1-mini and o1-preview on middle-school, high-school, college-entrance and MATH benchmarks
Same report: 90% of o1-mini's score on OMNI-MATH, 83% on AIME

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

mathematical-reasoningmath-educationcompetition-math

Trust score — 6/10

/10 trustMetrics-checked

Scored 6/10 (TIER metrics, metrics-v1). Evidence: Moonshot AI announcement (Nov 16 2024) via TMTPost and AInvest with comparative claims vs o1-mini/o1-preview; rolled out on the Kimi platform. Strengths: notable early RL-scaling milestone from a major Chinese lab. Docked for: (1) closed weights and limited public technical detail — no paper or model card; (2) all benchmark claims are vendor-reported; (3) superseded within Moonshot's own lineup. License: proprietary.

How the trust score is built

Full methodology →

Adoption · stars/downloads, log-scaled vs peers: 6.0/10
Maintenance · recency, releases, issue load: 5.0/10
Openness · license + docs: 2.0/10
Manual review · sanity read + red-flag check: 6.0/10

· Moonshot AI; announced 2024-11-16
· Beat o1-mini/o1-preview on school benchmarks; 83% of o1-mini on AIME
· No paper or model card; benchmarks vendor-reported